Settlement Manipulation in Prediction Markets

Battle Count: 50

Highly relevant to quantitative trading in several dimensions: (1) directly addresses market manipulation detection and its impact on price discovery, which is critical for any trading strategy relying on market prices; (2) the Glosten-Milgrom framework and adverse selection modeling are foundational to market-making and execution algorithms; (3) the finding that manipulation makes spot markets more liquid but less informative has direct implications for liquidity provision strategies; (4) the identification of manipulation windows (final 10 seconds, NTM cycles, thin-liquidity hours) provides actionable intelligence for trading systems; (5) the wealth transfer from retail to manipulators (93% of losses on retail) informs risk management for retail-facing trading platforms; (6) the contract horizon as a design lever is relevant for any firm developing or trading binary/derivative products; (7) the on-chain data methodology provides a template for manipulation detection in other markets.

The theoretical model involves solving a sequential game with multiple stages, fixed-point arguments for the manipulator's push probability, and comparative statics across parameters (σ, π, ε). The empirical analysis requires processing 60M+ on-chain trade records, matching them to cycle-level aggregates, constructing PushIntensity metrics, classifying NTM/FFM cycles, running multiple OLS specifications with fixed effects and clustered standard errors, and performing wallet-level PnL attribution across 243,155 traders. The difference-in-differences design for the fifteen-minute contract adds further complexity. However, all methods are standard in market microstructure literature, and the data is publicly available.

Highly reproducible due to public on-chain data (Polymarket trades on Polygon blockchain, Binance spot data publicly available). The paper provides detailed construction of all variables, regression specifications, and identification strategies. However, the theoretical model requires careful parameter calibration, and the manipulator identification relies on ad hoc thresholds (≥5 manipulated cycles, ≥$2,000 PnL), though sensitivity analysis (Table A5) shows robustness. The Chainlink oracle resolution data availability is limited to a subset of markets (Mar 16 - Apr 10, 2026 window). No code repository is mentioned.